Grok 5: Elon Musk Is Building Something Insane
Link to our newsletter: https://bitbiased.ai/ Grok 4.7 was supposed to launch on September 11th, 2026. It didn’t. Just two days later, Elon Musk had already shifted the conversation to Grok 4.8 — and an even bigger model behind it: Grok 5. xAI says Grok 5 is planned before the end of 2026, and Musk has claimed it will train on the entire SpaceX data corpus, covering more than 25 years of engineering records and employee output. But other eye-catching numbers surrounding the model — including a possible six trillion parameters and Musk’s earlier claim of a roughly 10% chance of AGI — are far less certain. That distinction matters, because Grok 5 is already accumulating enormous expectations despite having no public architecture, context window, benchmark preview, or confirmed training-compute specification. The hardware behind it, however, is much more concrete. Colossus II is xAI’s massive Blackwell-generation expansion in Memphis. Plans discussed in the video point toward roughly 440,000 GB200 and GB300 GPUs in Colossus II alone. Combined with the existing campus, independent analysis puts the operation at roughly 555,000 GPUs and close to 2 gigawatts of power draw. Then there’s the cost. GPU hardware alone is estimated at around $18 billion, while the complete infrastructure could potentially reach $35–58 billion once buildings, cooling, battery storage, and power systems are included. So what does that amount of compute actually buy xAI? Grok 5 could make a substantial jump in reasoning depth and context handling. Coding is less straightforward. Grok 4.6 trails its closest contemporaneous rival on Stanford’s APEX-SWE benchmark in the figures examined here, suggesting that simply adding parameters and compute may not be enough. Better software-engineering training data and reinforcement learning could matter just as much. There’s also a multimodal catch. Grok currently relies on separate systems for capabilities such as image generation and voice rather than functioning as one unified model that natively handles every modality. Nothing publicly available confirms that Grok 5 changes that architecture. And the competition isn’t standing still. This video looks at where Grok 5 could realistically land against OpenAI’s GPT-6 Astra, Google Gemini 3.1 Pro, and Anthropic’s latest Claude models — and why increasingly narrow benchmark gaps between frontier labs make claims of a dramatic category-defining lead harder to justify. Finally, there’s Musk’s AGI framing. More compute can produce a significantly more capable model, but bigger context windows, stronger benchmarks, and longer reasoning chains do not automatically give a system autonomous learning, independent goal-setting, or self-improvement. The most remarkable part of Grok 5 may therefore be the part that already exists: xAI has built an extraordinary amount of physical infrastructure around its next generation of models. Whether Grok 5 turns that investment into a genuine frontier-model breakthrough is still an open question. CHAPTERS 00:00 Grok 5 Is Already Surrounded by Huge Claims 01:42 The Grok 4.7 Disappearing Act 03:12 What xAI Has Actually Confirmed About Grok 5 04:29 Colossus II: The Machine That Would Train It 06:02 What That Compute Could Actually Buy 07:51 The Multimodal Reality Check 08:40 Where This Actually Lands Against the Competition 09:47 Is This Actually a Step Toward AGI 10:57 The Bottom Line #grok5 #xai #elonmusk #artificialintelligence #agi
